You want to use AI on your phone, but you’re stuck in a dead zone. Typically, that won’t work, as those large language models require an internet connection. Aah, but wait. Google has cooked up an app designed to surmount that barrier.
Available now for Android and coming soon to iOS, the experimental Google AI Edge Gallery will download the necessary models and files required to use an AI. You can then carry out the usual tasks, such as asking questions and finding information, even if you’re offline.
Also: Your Google Gemini assistant is getting 8 useful features – here’s the update log
“The Google AI Edge Gallery is an experimental app that puts the power of cutting-edge Generative AI models directly into your hands, running entirely on your Android (available now) and iOS (coming soon) devices,” Google said on the app’s GitHub page. “Dive into a world of creative and practical AI use cases, all running locally, without needing an internet connection once the model is loaded. Experiment with different models, chat, ask questions with images, explore prompts, and more.”
How does this work?
The setup process is awkward, but doable. Since it’s experimental, Google AI Edge Gallery isn’t available through Google Play or other official app stores. Instead, you’ll have to install it through an APK file. For that, you’ll need to tweak the settings on your Android device to allow the installation of unknown apps. Plus, you must have an account with Hugging Face, through which you download the LLM versions that will run locally.
“On your device, head to Settings and select the setting for Apps. Depending on your phone and OS version, either tap the option at the bottom for “Special app access” or tap the three-dot icon at the top and select “Special access. At the next screen, tap the option for “Install unknown apps” and then turn on the switch for Chrome to allow from that source.
Also: How to try Veo 3, Google’s AI video generator that’s going viral on the internet
To grab the app, head to the GitHub page through Chrome on your Android device. Swipe down the page past the screenshots and tap the link for “Download the App: Grab the latest APK.” If you receive a notification telling you that the file might be harmful, tap the option to download it anyway.
Tap the Open link, and then tap Install. After the installation, tap Open to launch the app. For future use, you’ll also find the app’s icon in your app library.
To get you started, the app offers three example tasks to try — Ask Image, AI Chat, and Prompt Lab. Tap any of them. You’ll then have to choose one or more of the LLMs to download and use. Here, the app offers different versions of Google’s Gemma LLM, a lightweight and open model similar to Gemini, but able to run locally. Tap the double arrow next to any version to learn more about it. Otherwise, select one and tap “Download & Try.”
To proceed, you’ll need to create or sign into a Hugging Face account. After logging in, you’ll have to agree to the license terms. That process can be clumsy, but don’t give up. The version you choose is downloaded and installed on your device. A couple of them are several gigabytes in size, so be patient. You can then download other versions if you wish.
Also: 3 ways Google’s AI Mode will change how you shop online – if it works
When done, return to the main Google AI Edge Gallery screen and tap one of the three examples. Select the downloaded model you wish to use, and then tap Try. Choose AI Chat, for example, and then type or speak your prompt. In response, the AI will generate its response. I asked it when and where pizza was invented, and the AI delivered a detailed response.
Next, you may want to try Ask Image. Here, you can upload an image from your device or take a new photo and then ask questions about it. However, I ran into some hiccups when I tried this.
My experience
I took a photo of a model of the USS Enterprise from the original Star Trek and asked the AI how many crew members served aboard. The AI correctly identified the ship, but kept coming up with wildly inaccurate numbers for the crew count, like 22 and 10 (the actual number is around 430). After I called it out on each mistake, it apologized and said that it was still under development and still learning. I even tried different models, but it continued to give me wrong answers.
Also: How to run a local LLM as a browser-based AI with this free extension
I also showed it a picture of the cover of Amazing Spider-Man #1 and asked it to identify the comic book. On the first attempt, it informed me that it was Amazing Spider-Man #2. When I told it that answer was wrong, it then served up the right information.
Beyond the hallucinations, the app itself crashed a few times when I was working with the Ask Image prompt. Of course, AI Edge Gallery is labeled an experiment for a reason. Don’t expect a stable performance or accurate answers across the board. But if Google is able to fine-tune the app, then it could pave the way for AIs that can run effectively whether you’re online or not.
Want more stories about AI? Sign up for Innovation, our weekly newsletter.
Leave a Reply